cohort-analysislisted

> **Provenance.** Vendored from `phuryn/pm-skills@cohort-analysis` (MIT, Paweł Huryn / Product Compass — github.com/phuryn/pm-skills). Adapted into great-pm 2026-05-29 with attribution; reframed 2026-06-17 from a data-engineering recipe into a PM interpretation playbook. Host agent: analytics-analyst. # Cohort Analysis — interpretation playbook A cohort is a group of users bucketed by a shared starting event (signup month, first-purchase week, feature-launch exposure) and then tracked over the same elapsed time. The chart is the easy part. **The skill is the judgement: deciding what is a real difference between cohorts, what is noise, and what is an artefact of how you drew the cohort.** This skill is a decision playbook, not a pandas recipe — you produce an insight and a recommendation, not a script (generate code only if the human explicitly asks). ## 1. Define the cohort — and freeze the definition Before any number, pin down three things and write them down: - **Grouping event** — what makes someone a member? (signup week, first paid transaction, first exposure to feature X). One event, stated once. - **Elapsed-time axis** — week 0, 1, 2… measured from the grouping event, NOT calendar time. A user who signed up in March and one who signed up in May are both "week 4" at four weeks after their own signup. - **The action that counts as "retained"** — opened the app? completed a core action? paid? "Retention" with no defined action is meaningless. If you cannot